diff --git a/backends/vulkan/runtime/VulkanBackend.cpp b/backends/vulkan/runtime/VulkanBackend.cpp
index 14cc7df02a1..5093d61d2f7 100644
--- a/backends/vulkan/runtime/VulkanBackend.cpp
+++ b/backends/vulkan/runtime/VulkanBackend.cpp
@@ -621,6 +621,10 @@ class VulkanBackend final : public ::executorch::runtime::BackendInterface {
   void destroy(DelegateHandle* handle) const override {
     if (handle != nullptr) {
       ComputeGraph* compute_graph = static_cast<ComputeGraph*>(handle);
+      compute_graph->context()
+          ->adapter_ptr()
+          ->compute_pipeline_cache()
+          .save_cache();
       // ComputeGraph is not trivially destructible. Since
       // this was constructed manually in init(), we must destroy it manually
       // here.
diff --git a/backends/vulkan/runtime/vk_api/Pipeline.cpp b/backends/vulkan/runtime/vk_api/Pipeline.cpp
index 07fc313984a..49bbf083359 100644
--- a/backends/vulkan/runtime/vk_api/Pipeline.cpp
+++ b/backends/vulkan/runtime/vk_api/Pipeline.cpp
@@ -406,8 +406,6 @@ ComputePipelineCache::~ComputePipelineCache() {
     return;
   }
 
-  save_cache();
-
   vkDestroyPipelineCache(device_, pipeline_cache_, nullptr);
   pipeline_cache_ = VK_NULL_HANDLE;
 }
diff --git a/backends/vulkan/runtime/vk_api/Pipeline.h b/backends/vulkan/runtime/vk_api/Pipeline.h
index 8765156deb2..4f42a9bf6bb 100644
--- a/backends/vulkan/runtime/vk_api/Pipeline.h
+++ b/backends/vulkan/runtime/vk_api/Pipeline.h
@@ -269,9 +269,10 @@ class ComputePipelineCache final {
     }
   };
 
+  void save_cache();
+
  private:
   std::vector<char> load_cache();
-  void save_cache();
 
   // Multiple threads could potentially be adding entries into the cache, so use
   // a mutex to manage access